AI's Invisible Hand: How Agents Dictate Tech Stacks and Drive Open-Source Evolution with 'Patch MD'

Recent analyses of AI language model behavior indicate a significant shift in how developers select and interact with software tools. When unsteered, models like Anthropic’s Claude Code frequently prefer building custom solutions from scratch but also converge on a consistent set of popular third-party tools, including Vercel, PostgreSQL, Stripe, Tailwind, ShadCN UI, PNPM, GitHub Actions, Sentry, Resend, and Zustand. Notably, Zustand now exceeds Redux in weekly npm installs, reflecting this trend. Newer AI models, particularly later Opus versions, demonstrate a greater willingness to recommend emerging tools, signifying a dynamic influence on market adoption.

Comparison between Claude Code and OpenAI’s Codex reveals divergent preferences; for instance, Claude may incorrectly recommend Bun for Next.js 14 runtime despite compatibility issues, while OpenAI models offer more nuanced and source-backed guidance. Claude also exhibits a consistent avoidance of Cloudflare services across various recommendations. These models are increasingly adept at processing documentation, allowing them to integrate and recommend new or private tools effectively. This evolving landscape underpins what Mitchell Hashimoto terms the ‘building block economy,’ where libraries (e.g., Lib Ghosty) are experiencing faster adoption than their full application counterparts (Ghosty), underscoring a preference for modular components over monolithic applications.

In response to this, a new paradigm for open-source business models is emerging. While open sourcing presents challenges like security vulnerabilities and self-hosting risks, it also fosters outsourced R&D and community-driven feature development. The T3 Code project, an open-source AI coding agent wrapper, exemplifies this, with 10% of its 16,000 weekly users creating custom forks. To manage this deep customization and ensure maintainability, the ‘patch MD’ concept is proposed: a standardized text file defining software modifications. This approach aims to enable self-forking, self-customizing, and self-healing software, allowing AI agents to resolve merge conflicts during updates and empowering users to tailor applications without constant manual intervention, heralding a future of highly malleable and community-driven software development.